Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
import { type CSSProperties, useEffect, useId, useLayoutEffect, useRef, useState } from 'react'
import { createPortal } from 'react-dom'
import { Button } from '@components/Common'
import type { PluginConsentModalProps } from './types'
/**
* The plugin's live rectangle in the renderer is the `.sandbox-ui-embed-slot`
* div that the native WebContentsView floats above (see SandboxUiPage). While
* this prompt is up the native view is hidden, but the slot stays laid out — so
* it is the correct box to center the prompt over, rather than the whole
* desktop window. Returns null when no plugin embed is mounted, which lets the
* prompt fall back to full-window centering.
*/
function readEmbedSlotRect(): DOMRect | null {
const el = document.querySelector('.sandbox-ui-embed-slot')
if (!el) return null
const rect = el.getBoundingClientRect()
return rect.width > 0 && rect.height > 0 ? rect : null
}
/**
* The plugin permission prompt (spec §9.2, §9.4).
*
* Rendered by the TRUSTED renderer while the plugin's own `WebContentsView` is
* hidden by the main process — a WebContentsView paints above renderer DOM
* regardless of z-index, so hiding it is the only way to guarantee the plugin
* can neither fake this modal nor paint over it.
*
* Every string here except `pluginTitle` comes from the host's own capability
* catalog. The plugin cannot influence a single word of what the user reads,
* and it cannot mark a row "required": that pressure belongs in the plugin's
* own UI after a partial grant, where the user can see who is asking.
*
* Rows arrive pre-sorted most-sensitive-first and start checked; unchecking is
* how a user grants part of a request.
*/
export function PluginConsentModal({ request, onResolve }: PluginConsentModalProps) {
const titleId = useId()
const allowButtonRef = useRef<HTMLButtonElement | null>(null)
const [checked, setChecked] = useState<Set<string>>(
() => new Set(request.rows.map(row => row.capability))
)
useEffect(() => {
setChecked(new Set(request.rows.map(row => row.capability)))
}, [request.promptId, request.rows])
useEffect(() => {
allowButtonRef.current?.focus()
}, [request.promptId])
// Track the plugin embed's rectangle so the prompt centers over the plugin
// UI, not the entire desktop app. Re-measure on the same signals that move
// the embed slot (resize, scroll, sidebar/layout shifts).
const [slotRect, setSlotRect] = useState<DOMRect | null>(null)
useLayoutEffect(() => {
const measure = (): void => setSlotRect(readEmbedSlotRect())
measure()
const slot = document.querySelector('.sandbox-ui-embed-slot')
const observer = slot ? new ResizeObserver(measure) : null
if (slot) observer?.observe(slot)
window.addEventListener('resize', measure)
window.addEventListener('scroll', measure, true)
return () => {
observer?.disconnect()
window.removeEventListener('resize', measure)
window.removeEventListener('scroll', measure, true)
}
}, [request.promptId])
useEffect(() => {
const handleKeyDown = (event: KeyboardEvent) => {
// Escape denies everything — a prompt nobody answered is not consent.
if (event.key === 'Escape') onResolve(request.promptId, [])
}
window.addEventListener('keydown', handleKeyDown)
return () => window.removeEventListener('keydown', handleKeyDown)
}, [onResolve, request.promptId])
const toggle = (capability: string) => {
setChecked(previous => {
const next = new Set(previous)
if (next.has(capability)) next.delete(capability)
else next.add(capability)
return next
})
}
// Confine the fixed backdrop to the plugin's rectangle so `place-items:
// center` centers the dialog over the plugin UI. Setting right/bottom to auto
// overrides the stylesheet's `inset: 0`. No slot → undefined → full-window.
const backdropStyle: CSSProperties | undefined = slotRect
? {
top: slotRect.top,
left: slotRect.left,
width: slotRect.width,
height: slotRect.height,
right: 'auto',
bottom: 'auto',
}
: undefined
return createPortal(
<div
className="confirm-dialog-backdrop da-plugin-consent-backdrop"
role="presentation"
style={backdropStyle}
>
{/* Deliberately NOT aria-modal: the backdrop is scoped to the plugin's
embed rectangle (see backdropStyle above), so the surrounding trusted
app chrome stays visible and interactive by design. Claiming full-window
modality here would misdescribe the DOM to assistive tech. The plugin
cannot exploit the un-dimmed chrome — main hides the plugin's own
WebContentsView while this prompt is up. */}
<div className="confirm-dialog da-plugin-consent" role="dialog" aria-labelledby={titleId}>
<h3 id={titleId}>{request.pluginTitle} wants access to your information</h3>
{request.priorPromptCount > 0 ? (
<p className="muted da-plugin-consent__count">
This plugin has asked for permissions {request.priorPromptCount + 1} times this session.
</p>
) : null}
<ul className="da-plugin-consent__rows">
{request.rows.map(row => (
<li key={row.capability} className="da-plugin-consent__row">
<label className="da-plugin-consent__label">
<input
type="checkbox"
checked={checked.has(row.capability)}
onChange={() => toggle(row.capability)}
/>
<span>
<span className="da-plugin-consent__row-title">{row.title}</span>
<span className="da-plugin-consent__row-detail muted">{row.dataDescription}</span>
</span>
</label>
</li>
))}
</ul>
<p className="muted da-plugin-consent__footer">
You can change this any time in Settings → Plugin permissions.
</p>
<div className="confirm-dialog-actions">
<Button
color="neutral"
onClick={() => onResolve(request.promptId, [])}
size="sm"
variant="ghost"
>
Deny
</Button>
<Button
ref={allowButtonRef}
color="primary"
disabled={checked.size === 0}
onClick={() => onResolve(request.promptId, [...checked])}
size="sm"
>
Allow
</Button>
</div>
</div>
</div>,
document.body
)
}